Liquid AI releases 230M-parameter LFM2.5 model

Liquid AI announced LFM2.5-230M, its smallest model to date. Base and post-trained checkpoints are on Hugging Face. At launch it supports llama.cpp (GGUF), MLX, vLLM, SGLang and ONNX.

The model uses the LFM2 architecture and was pre-trained on 19T tokens, including a phase that extended context to 32K. Post-training happened in three stages: supervised fine-tuning with distillation from LFM2.5-350M, then direct preference optimization, then multi-domain reinforcement learning. Liquid says it kept the recipe light so developers can still fine-tune the model for their own tasks.

Liquid reports decode speeds of 213 tokens/s on a Samsung Galaxy S25 Ultra and 42 tokens/s on a Raspberry Pi 5. It says the model has the highest prefill and decode throughput and the smallest memory footprint in its class on both devices.

On Liquid's ten-benchmark suite, the 230M model often beats Granite 4.0 350M variants and Gemma 3 1B. Examples include IFEval (71.71), BFCLv3 (43.26) and CaseReportBench (22.51). The results are mixed, though. Qwen3.5-0.8B scores far higher on MMLU-Pro (37.42 vs 20.25). The new model also scores just 5.26 on τ²-Bench Telecom, below most of the models it was compared with, including Liquid's own LFM2.5-350M at 18.86. Liquid advises against using it for advanced math, code generation or creative writing.

In a demo, Liquid fine-tuned the model to run entirely on the NVIDIA Jetson Orin inside a Unitree G1 humanoid robot. There it breaks a natural-language instruction into a sequence of calls to low-level skills from NVIDIA's SONIC framework. The company calls the behaviors "deliberately simple" at this stage.

Liquid also mentions an internal GPU inference stack with lower latency than small models served on SGLang. It did not publish figures for it.